HyphaeDB: Agent-Native Memory for Multi-Agent AI Systems

This paper presents HyphaeDB, a groundbreaking memory infrastructure designed specifically for multi-agent AI systems. Unlike traditional vector databases that treat memory as passive storage, HyphaeDB re-envisions the Hierarchical Navigable Small World (HNSW) graph—a core data structure in modern vector databases—not merely as a search optimization, but as an active communication network. In this system, individual AI agents are represented as nodes within a vector space, maintaining persistent positions. Knowledge within HyphaeDB propagates dynamically between agents via a gossip protocol, attenuated by energy-based mechanisms through the graph's neighbor structure. This dynamic propagation, combined with local interaction rules and the topological structure, facilitates the emergence of complex behaviors such as the detection of contradictions, the crystallization of patterns, and the formation of consensus among agents. The architecture is built upon primitives like knowledge nodes, topology edges, and memory diffs, and includes a multi-layer abstraction hierarchy. A reference implementation using PostgreSQL with pgvector is provided, demonstrating its application in a multi-agent software engineering methodology called Swarm-Driven Development.